Real-Estate Asset Management Company: Information Consolidation and AI-Ready Platform Proposal
Commissioned Work | Forward Deployed Engineer
Context and Problem
Contract and ledger information was scattered across personal Excel files and the like, so work and inquiries stalled whenever the person in charge was away.
My Role
Owned the requirements document and the architecture selection: merged the team's three drafts (core database and two business areas) into requirements v1.0 and addressed the review comments. The engagement covered the proposal and requirements definition.
Acceptance Criteria
The requirements document turns "remove the dependency on one person" into checkable acceptance criteria (first drafted by a teammate).
- Someone other than the person in charge can explain each case's progress and missing documents.
- The system reproduces the same amounts as the current Excel files, planned to be checked against past months of real data at three levels: billee, item, and line.
How the Architecture Was Chosen
Priorities, in order: usable by AI agents (AI Agent Ready), able to grow in stages without a rebuild, fast to deliver and a fit for the team's skills, and low operating effort and cost. The core was fixed so it runs on any cloud: object storage for originals, PostgreSQL for both the ledger and analytics, dbt Core for transformations and metric definitions, and MCP as the access point for AI.
Proposed Architecture
- Originals: received documents are stored unchanged and immutable so they can be reprocessed later.
- Ingestion: a multimodal LLM (swappable) extracts fields from documents, and a person approves them before they enter the ledger; low-confidence extractions go to human review.
- Database: since the data volume is small, the ledger and analytics share one PostgreSQL database, and a dedicated data warehouse was judged excessive. pgvector was kept as an option for future semantic search.
- Semantic layer: dbt Core as the first choice, with schema and metric definitions in code.
- Ledger data model: buildings and units, tenants and contracts, and contracts with charges and move-in/move-out events as parent-child relations, with one contract able to span several units. Deletes are logical, and every change is kept in history (designed by a teammate; I added the diagram and fields).
- AI-ready: five guarantees (immutable originals, schema and metrics as code, lineage from each original to each record, human review of low-confidence extractions, and a multi-LLM design), with permissions and audit logs from day one. The AI access point is designed as a read-only MCP server with audit logging. Building the agent itself was out of scope.
- Staged growth: the ledger schema is frozen as v1 at release, and later changes are additive only.
Rejected Alternative
A BaaS would have been fast to build on, but it was rejected for its limits on permission integration, batch processing, and auditing.